Evaluating feature attribution methods in the image domain
Open Access
- 24 May 2024
- journal article
- research article
- Published by Springer Nature in Machine Learning
- Vol. 113 (9) , 6019-6064
Abstract
Feature attribution maps are a popular approach to highlight the most important pixels in an image for a given prediction of a model. Despite a recent growth in popularity and available methods, the objective evaluation of such attribution maps remains an open problem. Building on previous work in this domain, we investigate existing quality metrics and propose new variants of metrics for the evaluation of attribution maps. We confirm a recent finding that different quality metrics seem to measure different underlying properties of attribution maps, and extend this finding to a larger selection of attribution methods, quality metrics, and datasets. We also find that metric results on one dataset do not necessarily generalize to other datasets, and methods with desirable theoretical properties do not necessarily outperform computationally cheaper alternatives in practice. Based on these findings, we propose a general benchmarking approach to help guide the selection of attribution methods for a given use case. Implementations of attribution metrics and our experiments are available online (https://github.com/arnegevaert/benchmark-general-imaging). Graphical abstractKeywords
All Related Versions
Funding Information
- Onderzoeksprogramma Artificiele Intelligentie (AI) Vlaanderen
- Bijzonder Onderzoeksfonds UGent (01D13919)
This publication has 35 references indexed in Scilit:
- Explainable artificial intelligence (XAI) in deep learning-based medical image analysisMedical Image Analysis, 2022
- Improving performance of deep learning models with axiomatic attribution priors and expected gradientsNature Machine Intelligence, 2021
- Multi-Objective Counterfactual Explanations2020
- Mastering the game of Go without human knowledgeNature, 2017
- Human-level control through deep reinforcement learningNature, 2015
- Microsoft COCO: Common Objects in Context2014
- Visualizing and Understanding Convolutional Networks2014
- A survey on feature selection methodsComputers and Electrical Engineering, 2014
- Causality2009
- Gradient-based learning applied to document recognitionProceedings of the IEEE, 1998